Download Algebraic D-groups and differential Galois theory

PACIFIC JOURNAL OF MATHEMATICS
Vol. 216, No. 2, 2004
ALGEBRAIC D-GROUPS AND DIFFERENTIAL GALOIS
THEORY
Anand Pillay
We discuss various relationships between the algebraic Dgroups of Buium, 1992, and differential Galois theory. In
the first part we give another exposition of our general differential Galois theory, which is somewhat more explicit than
Pillay, 1998, and where generalized logarithmic derivatives on
algebraic groups play a central role. In the second part we
prove some results with a “constrained Galois cohomological
flavor”. For example, if G and H are connected algebraic
D-groups over an algebraically closed differential field F , and
G and H are isomorphic over some differential field extension of F , then they are isomorphic over some Picard–Vessiot
extension of F . Suitable generalizations to isomorphisms of
algebraic D-varieties are also given.
1. Introduction
We work throughout with (differential) fields of characteristic zero. In
[8] the notion of a generalized differential Galois extension (or generalized
strongly normal extension) of a differential field was introduced, generalizing
Kolchin’s theory of strongly normal extensions, which in turn generalized
the Picard–Vessiot theory. The idea was to systematically replace algebraic groups over the constants by “finite-dimensional differential algebraic
groups”, to obtain new classes of extensions of differential fields with a good
Galois theory. This idea (almost obvious from the model-theoretic point of
view) was implicit in Poizat [11] who gave a model-theoretic treatment of
the strongly normal theory. However the “correct” definition of a generalized
differential Galois extension needed some additional fine-tuning. Nevertheless, our exposition of this general theory in [8] was overly model-theoretic,
and possibly remained somewhat obscure to differential algebraists. We try
to remedy this in the current paper by concentrating on the differential equations that have a good Galois theory, very much in the spirit of Section 7,
Chapter IV of Kolchin’s book [4]. The key notion is that of a generalized
logarithmic derivative on an algebraic group G over a differential field K
(a certain kind of differential rational map from G to its Lie algebra). We
343
344
ANAND PILLAY
will see that such a generalized logarithmic derivative is essentially equivalent to an algebraic D-group structure on G (in the sense of Buium [3]).
Our resulting exposition of the generalized differential Galois theory will be
equivalent to that in [8] when the base field K is algebraically closed. The
general situation (K not necessarily algebraically closed) can be treated using analogues of the V -primitives from [4, IV.10], and we leave the details
to others.
Let me now say a little more about the generalized logarithmic derivatives,
and how they tie up with the Picard–Vessiot/strongly normal theory. Let us
fix a differential field K, and assume for now that the field CK of constants
of K is algebraically closed. A linear differential equation over K, in vector
form, is ∂y = Ay, where y is a n × 1 column vector of unknowns and A is
an n × n matrix over K. Looking for a fundamental matrix of solutions,
one is led to the equation on GLn : ∂Y = AY , where Y is a n × n matrix
of unknowns ranging over GLn , which we can write as ∂(Y )Y −1 = A. Now
the map Y → ∂(Y )Y −1 is the classical logarithmic derivative, a first-order
differential crossed homomorphism from GLn into its Lie algebra, which is
surjective when viewed in a differentially closed overfield of K. A Picard–
Vessiot extension of K for the original equation is then a differential field
extension L = K(g), where g ∈ GLn is a solution of ∂(Y )Y −1 = A, and
CL = CK . Such an extension exists, and is unique up to K-isomorphism.
The group of (differential) automorphisms of L over K has the structure of
an algebraic subgroup of GLn (CK ), and there is a Galois correspondence.
In place of GLn one can consider an arbitrary algebraic group G defined
over K (not necessarily linear and not necessarily defined over the constants
of K). By a generalized logarithmic derivative on G we will mean a first-order
differential rational crossed homomorphism µ from G to L(G), defined over
K, such that µ is geometrically surjective (that is, surjective when viewed in
a differentially closed overfield) and such that Ker (µ), a finite-dimensional
differential algebraic subgroup of G, is Zariski-dense in G. The analogue of
a linear differential equation over K will then be an equation
(∗)
µ(x) = a,
where x ranges over G and a ∈ L(G)(K).
Under an additional technical condition on the data, analogous to the
requirement that the field of constants of K be algebraically closed, we can
define the notion of a differential Galois extension L of K for the equation
(∗), prove its existence and uniqueness, identify the Galois group, and obtain
a Galois correspondence. In the case where G is defined over the constant
field CK and µ is the standard logarithmic derivative of Kolchin, we recover
Kolchin’s strongly normal extensions (see Theorem 6, Section 7, Chapter IV
of [4]).
ALGEBRAIC D-GROUPS
345
For G an algebraic group over the differential field K, an algebraic Dgroup structure on G is precisely an extension of the derivation ∂ on K to
a derivation on the structure sheaf of G, respecting the group operation.
Algebraic D-groups belong entirely to algebraic geometry, and Buium [3]
points out that there is an equivalence of categories between the category of
algebraic D-groups and the category of ∂0 -groups, finite-dimensional differential algebraic groups. The latter category belongs to Kolchin’s differential
algebraic geometry. On the other hand, there is essentially a one-to-one
correspondence between algebraic D-group structures on G and generalized
logarithmic derivatives on G. So our general differential Galois theory is in
a sense subsumed by the very concept of an algebraic D-group.
Details of the above will be given in Sections 2 and 3, including a “Tannakian” approach and an examination of different manifestations of the differential Galois group.
In Section 4 we will give another relation between algebraic D-groups and
the Picard–Vessiot theory: if two algebraic D-groups over an algebraically
closed differential field are isomorphic (as D-groups) over some differential
field extension of K, then such an isomorphism can be found defined over a
Picard–Vessiot extension of K. This uses Kolchin’s differential Lie algebra,
and strengthens the somewhat artificial results from [9]. We will also give
some related results on isomorphisms between algebraic D-varieties, using
differential jets (higher-dimensional versions of differential tangent spaces).
2. Algebraic D-groups
We will briefly describe the algebraic D-groups of Buium in a matter suitable
for our purposes. We also introduce the generalized logarithmic derivative
on an algebraic group induced by a given D-group structure. We refer to
[6] for a discussion of related themes.
Let us fix an ordinary differential field (K, ∂). For convenience we also
give ourselves a differential field extension (U, ∂) of (K, ∂) that is “universal”
with respect to (K, ∂). Namely, U has cardinality κ > |K|, and for any
differential subfield F < U of cardinality < κ and differential extension L
of K of cardinality ≤ κ there is an embedding (as differential fields) of L
into U over K. CK denotes the field of constants of K, and C the field of
constants of U.
We begin by recalling the tangent bundle of an algebraic variety or group
over K (in which the derivation on K plays no role).
Let X be an algebraic variety over K (maybe reducible). The tangent
bundle T (X) of X is another algebraic variety over K, with a canonical
surjective
morphism π (over K) to X, and is defined locally by equations:
∂P/∂x
(x1 , . . . , xn )vi = 0 for P polynomials over K generating the ideal
i
i
of X over K. If X = G is an algebraic group over K, then T (G) has the
346
ANAND PILLAY
structure of an algebraic group over K such that the canonical projection to
G is an (algebraic) group homomorphism. The group operation on T (G) is
obtained by differentiating the group operation of G. That is, if f (−, −) is
the group operation on G and (g, u) and (h, v) are in T (G) then the product
(g, u) · (h, v) equals (g · h, dfg,h (u, v)). Note that if λg , ρg denote left and
right multiplication by g in G, then we have:
(∗)
In T (G),
(g, u) · (h, v) = (g · h, d(λg )h (v) + d(ρh )g (u)).
We will denote T (G)e , the tangent space of G at the identity, by L(G)
(for the Lie algebra of G). Then L(G), with its usual vector space group
structure, is a normal subgroup of T (G), and we have the exact sequence
0 → L(G) → T (G) → G → {e} of algebraic groups (over K). We denote by
i : L(G) → T (G) the natural inclusion, and by π : T (G) → G the canonical
surjection above.
Note that G acts on L(G) (denoted (g, a) → ag ) by the adjoint map
(differentiating conjugation by g ∈ G at the identity). And this “coincides”
with the action of T (G) on the normal subgroup L(G) by conjugation: for
a ∈ L(G) and g ∈ G, ag = xax−1 for any x ∈ T (G) such that π(x) = g.
For a ∈ L(G), we let la and ra denote the left and right invariant vector
fields on G determined by a. Namely for g ∈ G, la (g) = d(λg )e (a) and
ra (g) = d(ρg )e (a).
A K-rational splitting of T (G) as a semidirect product of G and L(G) is
given by either of the equivalent pieces of data:
(a) a K-rational homomorphic section s : G → T (G) (that is π ◦ s = id);
(b) a K-rational crossed homomorphism h from T (G) onto L(G) such that
h◦i = id. (By definition α : T (G) → L(G) is a crossed homomorphism
if α(xy) = α(x) + xα(y)x−1 .)
Note that the set of these K-rational splittings has the structure of a
commutative group. For example, using the data in (a), if s1 , s2 are Krational homomorphic sections of the tangent bundle, and si (g) = (g, ui ) for
i = 1, 2 then (s1 + s2 )(g) = (g, u1 , +u2 ). The identity element is just the
0-section s0 (g) = (g, 0), which is a K-rational homomorphic section. Let
PK (G) denote this commutative group of K-rational splittings of T (G). We
denote the crossed homomorphism from T (G) onto L(G) corresponding to
the identity of PK (G) by h0 : T (G) → L(G), and the crossed homomorphism
corresponding to the homomorphic section s by hs .
Remark 2.1.
−1
(i) h0 (g, u) = d(ρg )g (u).
(ii) More generally, if s is a K-rational homomorphic section of T (G) → G,
−1
then hs (g, u) = d(ρg )g (u − s(g)).
Proof. This follows directly from formula (∗) for multiplication in T (G). ALGEBRAIC D-GROUPS
347
Let us now bring in the differential structure. Assume for now that the
algebraic variety X is defined over CK the field of constants of K. Then it is
easy to see that if a ∈ X(U) then working in local coordinates with respect
to a given covering of X by affine varieties over CK , (a, ∂(a)) ∈ T (X). If
in addition X = G is an algebraic group defined over CK , then the map
∇ : G(U) → T (G)(U) taking g to (g, ∂(g)) is a group embedding. Let
lD : G(U) → L(G)(U) be defined by lD(g) = h0 (g, ∂(g)). Then lD is
a “differential rational” crossed homomorphism defined over K, which is
precisely Kolchin’s logarithmic derivative. The map lD depends on h0 , and
clearly any other h ∈ PK (G) gives rise to another differential rational crossed
homomorphism (over K) from G(U) onto L(G)(U).
Remark 2.2. Suppose G is defined over CK . Then:
(i) The (standard ) logarithmic derivative lD : G(U) → L(G)(U) is given
−1
by lD(g) = d(ρg )g (∂(g)).
(ii) lD is surjective.
(iii) Ker (lD) is precisely G(C).
Proof. (i) follows from Remark 2.1.
(ii) Let a ∈ L(G)(U). Then a determines the right invariant vector field
ra : G(U) → T (G)(U). As U is differentially closed, the main result of [7]
gives g ∈ G(U) such that ∂(g) = ra (g), hence (by (i)), lD(g) = a.
(iii) is obvious from (i).
Let us now work in a more general context, dropping our assumption
that the variety X is defined over CK . Then for a ∈ X(U), (a, ∂(a)) may
no longer be a point of T (X) but rather a point of another bundle τ (X)
over X, which we now describe. For P (x1 , . . . , xn ) a polynomial over K, let
P ∂ denote the polynomial obtained from P by applying ∂ to its coefficients.
(So P ∂ = 0 if P is over CK .) Then τ (X) is defined locally by equations:
∂P/∂xi (x1 , . . . , xn )vi + P ∂ (x1 , . . . , xn ) = 0,
i
where P again ranges over polynomials in the ideal of X over K. (That
is, these affine pieces fit together to give an algebraic variety τ (X) over K,
together with a canonical projection from τ (X) to X.) It is immediate that
(a, ∂a) ∈ τ (X) for a ∈ X(U). It is also immediate that for each a ∈ X,
τ (X)a is a principal homogeneous space for the tangent space T (X)a , where
the action is addition (with respect to local coordinates above). Moreover
this happens uniformly, making τ (X) a torsor for the tangent bundle T (X).
In any case, if X is defined over CK , then τ (X) coincides with T (X). By an
algebraic D-variety over K we mean a pair (X, s) such that X is an algebraic
variety over K and s : X → τ (X) is a regular section defined over K.
348
ANAND PILLAY
Now assume again that X = G is an algebraic group over K. Then
τ (G) has the structure of an algebraic group over K such that the canonical projection τ (G) → G is a homomorphism. In local coordinates, assuming that multiplication in G is given by the sequence of polynomials
f = (fi (x1 , . . . , xn , y1 , . . . , yn ))i=1,...,n over K, then for (g, u), (h, v) ∈ τ (G),
the product of (g, u) and (h, v) in τ (G) is given by
g · h, df(g,h) (u, v) + (f1∂ (g, h), . . . , fn∂ (g, h)) .
We again have the map ∇ : G(U) → τ (G)(U), given (in local coordinates)
by ∇(g) = (g, ∂(g)) and this is a group embedding.
So we have two K-algebraic groups T (G) and τ (G). Although these
need not be isomorphic as algebraic groups, they are “differential rationally”
isomorphic. The following is left to the reader.
Lemma 2.3. The map that takes (g, u) to (g, ∂(g) − u) is a (differential
rational) isomorphism of groups between τ (G)(U) and T (G)(U). Although
not necessarily rational, it is rational when restricted to the fibers over G.
In particular, the above map defines a K-rational isomorphism between the
vector groups τ (G)e and T (G)e = L(G).
Note that we have again an exact sequence
0 → τ (G)e → τ (G) → G → e
of algebraic groups over K, which by virtue of the (canonical) isomorphism
between τ (G)e and L(G) given by Lemma 2.3 can be rewritten as:
0 → L(G) → τ (G) → G → e.
Let us again write i for the (canonical) injection of L(G) in τ (G), and π for
the canonical surjection τ (G) → G. So if G is defined over CK this agrees
with our earlier notation: i : L(G) → T (G) and π : T (G) → G.
We can consider splittings (as algebraic groups over K) of τ (G) as a
semidirect product of G and L(G). Again each such splitting is determined
either by a K-rational homomorphic section s : G → τ (G), or a K-rational
crossed homomorphism h : τ (G) → L(G) such that h ◦ i = id on L(G). We
will write hs for the crossed homomorphism corresponding to the homomorphic section s, and give explicit formulas below.
In any case, we can now define an algebraic D-group.
Definition 2.4. Let G an algebraic group over K. Then an algebraic Dgroup structure on G over K is precisely a K-rational homomorphic section
s : G → τ (G). We write the corresponding algebraic D-group as (G, s).
Given an algebraic D-group (G, s) we obtain a generalized logarithmic
derivative that we call lDs , a crossed homomorphism (in the obvious sense)
from G(U) to L(G)(U): lDs = hs ◦ ∇. Here is the analogue to Remarks 2.1
and 2.2.
ALGEBRAIC D-GROUPS
349
Remark 2.5. Let (G, s) be an algebraic D-group over K.
(i)
(ii)
(iii)
(iv)
−1
For (g, u) ∈ τ (G), hs (g, u) = d(ρg )g (u − s(g)).
−1
For g ∈ G(U), lDs (g) = d(ρg )g (∂(g) − s(g)) ∈ L(G)(U).
lDs : G(U) → L(G)(U) is surjective.
Ker (lDs ) is precisely {g ∈ G(U) : ∂(g) = s(g)}, and is a Zariski-dense
subgroup of G(U).
Proof. (i) follows from the formula for multiplication in τ (G), and (ii) is an
immediate consequence of (i).
(iii) Let a ∈ L(G)(U). Again we obtain the right invariant vector field
ra : G(U) → T (G)(U). Then ra + s : G(U) → τ (G)(U) is also a rational
section of τ (G) → G. By [7] there is g ∈ G(U) such that ∂(g) = ra (g)+s(g),
hence lDs (g) = a by (i).
(iv) Ker (lDs ) is a subgroup of G(U) as lDs is a crossed homomorphism.
−1
As d(ρg )g is an isomorphism between T (G)g and T (G)e , we see that
Ker (lDs ) is as described in (iii). By [7] for any proper subvariety X of
G(U) there is g ∈ G(U) such that ∂(g) = s(g). Hence by (i) Ker (lDs ) is
Zariski dense in G.
In the context of Remark 2.5, we denote Ker (lDs ) by (G, s) . This is a
finite-dimensional differential algebraic group, or ∂0 -group in the sense of
[3], and is an object belonging to Kolchin’s differential algebraic geometry.
Just for the record, here are some key properties: (G, s) (U) is Zariski-dense
in G, the ∂0 -subvarieties of (G, s) are precisely of the form X ∩ (G, s) for
X a D-subvariety of (G, s), and for any other algebraic D-group (H, t) the
∂0 -homomorphisms between (G, s) and (H, t) are precisely those induced
by algebraic D-group homomorphisms between (G, s) and (H, t).
In any case, we have seen that an algebraic D-group structure (G, s) on
an algebraic group G over K determines a generalized logarithmic derivative lDs on G. We point out now, just for completeness, that conversely
any suitable differential rational crossed homomorphism map (over K) from
G(U) to L(G)(U) determines an algebraic D-group structure on G.
Some notation: Let X and Y be algebraic varieties defined over K. By
a first-order differential rational map h : X(U) → Y (U), defined over K,
we mean a map h from X(U) to Y (U) such that for, for each x ∈ X(U),
h(x) ∈ K(x, ∂(x)).
Lemma 2.6. Suppose that G is a connected algebraic group over K. Let
l : G(U) → L(G)(U) be a first-order differential rational crossed homomorphism, defined over K, which is surjective, and is such that Ker (l) is
Zariski-dense in G(U). Then, there are a unique K-rational automorphism
σ of L(G), and a unique algebraic D-group structure (G, s) on G (defined
over K), such that l = σ ◦ lDs .
350
ANAND PILLAY
Proof. By [7], for example, ∇(G(U)) is Zariski-dense in τ (G)(U). Define
l1 on ∇(G(U)) by l1 (x, ∂(x)) = l(x). So by Zariski-denseness, and the
properties of l, l1 extends uniquely to a K-rational surjective crossed homomorphism f from τ (G) to L(G).
By the Zariski-denseness of Ker (l) in G and the definition of f , we have
π(Ker (f )) = G. On the other hand, dim(τ (G)) = 2 dim(G) and dim(G) =
dim(L(G)). Hence dim(Ker (f )) = dim(G). Now π|Ker (f ) : Ker (f ) → G
is a group homomorphism so it has finite kernel. But this finite kernel is
a subgroup of the vector group τ (G)e so has to be trivial. It follows that
f |τ (G)e is an isomorphism (over K) with L(G). So there is a unique Kautomorphism σ of L(G) such that (σ ◦ f ) ◦ i is the identity on L(G). Put
f1 = σ ◦f . Then f1 gives an algebraic D-group structure (G, s) on G defined
over K. For g ∈ G(U), σ ◦ l(g) = f1 (g, ∂(g)) = lDs (g).
There is a natural notion of a D-morphism between algebraic D-varieties
(X, s) and (Y, t). First note that τ (−) is a functor, so if f : X → Y is a
morphism between the algebraic varieties X and Y with everything defined
over K then τ (f ) is a morphism, over K, between τ (X) and τ (Y ), again
defined over f . (If X, Y, f are defined over the constants, then τ (f ) is just
the differential of f .) In any case, a morphism between algebraic D-varieties
X and Y is by definition a morphism f of algebraic varieties, such that
t ◦ f = τ (f ) ◦ s.
In particular we extract the notion of an algebraic D-subvariety of (X, s):
so an algebraic subvariety Y of X will be the underlying variety of an algebraic D-subvariety of (X, s) if s|Y maps Y to τ (Y ).
A homomorphism of algebraic D-groups is a D-morphism that is also a
homomorphism of algebraic groups.
We call an algebraic D-group (G, s) isotrivial if it is isomorphic over U to
a trivial algebraic D-group (H, s0 ), where H is defined over C and s0 is the
0-section of T (G).
The interest of the category of algebraic D-groups is that there exist
nonisotrivial algebraic D-groups. If A is an abelian variety over U and
p : G → A is the universal extension of A by a vector group then G has a
D-group structure (defined over the field over which A is defined). Moreover
any such D-group structure on G is nonisotrivial if A is not isomorphic (as
an algebraic group) to an abelian variety defined over C. Given such a Dgroup structure (G, s) on G, p((G, s) ) < A is precisely the Manin kernel of
A. This example is worked out in detail in [6].
Let us repeat from [9] the discussion of a nonisotrivial algebraic D-group
structure on the commutative algebraic group Gm ×Ga . So let G = Gm ×Ga .
Then T (G) = τ (G) can be identified with {(x, y, u, v) : x = 0} with group
structure (x1 , y1 , u1 , v1 )·(x2 , y2 , u2 , v2 ) = (x1 x2 , y1 +y2 , u1 x2 +u2 x1 , v1 +v2 ).
ALGEBRAIC D-GROUPS
351
Let s : G → T (G) be the homomorphic section s(x, y) = (x, y, xy, 0). Then
(G, s) is an algebraic D-group known to be nonisotrivial.
Note that if g = (x, y) ∈ G, then the differential of multiplication by
g −1 at g takes (u, v) ∈ T (G)g to (u/x, v) ∈ L(G). Hence by 2.5, the
generalized logarithmic derivative lDs corresponding to s is: lDs (x, y) =
((∂(x) − xy)/x, ∂(y)) = (∂(x)/x − y, ∂(y)).
In particular (G, s) = Ker (lDs ) = {(x, y) ∈ G : ∂(x)/x = y, ∂(y) = 0},
which can be identified with the subgroup of Gm defined by the second-order
equation ∂(∂(x)/x) = 0.
3. Differential Galois theory
The conventions of the previous section are in force. In particular (K, ∂) is
a differential field of characteristic 0, and we may refer also to the universal
differential field extension (U, ∂) of K.
By a logarithmic differential equation over a differential field (K, ∂) we
mean something of the form
(∗)
lDs (x) = a,
where (G, s) is an algebraic D-group defined over K and a ∈ L(G)(K). (So
the indeterminate x ranges over G.) As remarked in the introduction, a
special case is the equation ∂(X) = AX, where X ranges over GLn and A
is an n × n matrix over K.
In order to give the right notion of a differential Galois extension of K
for (∗), we need to place a further restriction on K, G and s.
Definition 3.1. Suppose (G, s) is an algebraic D-group over K. We say
that (G, s) is K-large, if for every (maybe reducible) algebraic D-subvariety
X of G, which is defined over K, X(K) ∩ (G, s) is Zariski-dense in X.
Remark 3.2.
(i) The intuitive meaning of (G, s) being K-large is that (G, s) has enough
points with coordinates in K.
(ii) Suppose that G is defined over CK and that s = s0 , the 0-section of
T (G). Then, if CK is algebraically closed, (G, s) is K-large.
In the next remark, we refer to differential closures of K. A differential
closure of K is a differential field extension of K that embeds over K into
any differentially closed field containing K. The differential closure of K
is unique up to K-isomorphism, and is written as K̂. Kolchin calls K̂ the
constrained closure of K.
Remark 3.3. (G, s) is K-large if and only (G, s) (K) = (G, s) (K̂) for
some (any) differential closure K̂ of K.
352
ANAND PILLAY
Proof. Assume first that (G, s) (K) = (G, s) (K̂). Let X be a D-subvariety
of (G, s) defined over K. The irreducible components X1 , . . . , Xr of X are
defined over K so also K̂ and are also D-subvarieties of (G, s). By [7] for
example for any nonempty Zariski open subset of any Xi defined over K̂,
there is a ∈ U (K̂ such that ∂(a) = s(a). By our assumptions a ∈ (G, s) (K).
So X ∩ (G, s) (K) is Zariski-dense in X.
Conversely, suppose (G, s) is K-large. Let a ∈ (G, s) (K̂), and suppose
for a contradiction that a ∈
/ G(K). Now a is constrained over K in the sense
of Section 10, Chapter III of [4]. This means that there is some differential
polynomial P (x) over K such that P (a) = 0, and whenever b is a differential
specialization of a over K and P (b) = 0 then b is a “generic” specialization
of a over K (in model-theoretic language tp(b/K) = tp(a/K)). Let X be
the irreducible K-subvariety of G whose generic point is a. Then (as ∂(a) =
s(a)), X is a D-subvariety of (G, s). Moreover the differential specializations
of a over K are precisely those b such that b ∈ X and ∂(b) = s(b). Now
subject to the conditions “x ∈ X and ∂(x) = s(x)”, the condition P (x) = 0
is clearly equivalent to x ∈ Y for Y some proper subvariety of X defined
over K. By our assumptions there is b ∈ G(K) such that b ∈ X \ Y and
∂(b) = s(b). This is a contradiction.
We call a differential ring (R, ∂) simple if it has no proper nontrivial
differential ideals. We refer to [12] for a discussion of simple differential
rings.
Lemma 3.4. Suppose that R is a simple differential ring over K that is
finitely generated over K. Then R embeds over K into some differential
closure of K.
Proof. As R has no zero-divisors, R embeds in U over K. Let R = K[a]∂ be
differentially generated over K by the finite tuple a. Suppose that π(a) = b
is a differential specialization over K. Then π extends to a surjective ring
homomorphism π : R → K[b]∂ . The kernel is a differential ideal, so must be
trivial. Thus b is a generic specialization of a over K. It follows that a is
constrained over K so lives in some differential closure of K, as does R. We can now give the main definition.
Definition 3.5. Let (G, s) be a K-large algebraic D-group defined over K,
and lDs (x) = a be a logarithmic differential equation over K for (G, s). By
a differential Galois extension of K for the equation lDs = a we mean a
differential field extension L of K of the form K(α) for some solution α of
the equation such that K[α], the (differential) ring generated by K and the
coordinates of α, is a simple differential ring.
Lemma 3.6 (Existence and uniqueness of differential Galois extensions).
If (G, s) is a K-large algebraic D-group defined over K, and a ∈ L(G)(K),
ALGEBRAIC D-GROUPS
353
then there exists a differential Galois extension L of K for the equation
lDs (x) = a. Moreover, any two such extensions are isomorphic over K as
differential fields.
Proof. By Remark 2.5 (iii) there is a solution β ∈ G(U) of lDs (x) = a. Let α
be a maximal differential specialization of β over K. Then K[α] is a simple
differential ring and α is also a solution of lDs = a, so we get existence. Let
L = K(α).
Suppose L1 is another differential Galois extension of K for the equation,
generated by the solution γ say. By Lemma 3.4 we may assume that both L
and L1 are contained in some differential closure K̂ of K. By Remark 3.3,
(G, s) (K̂) = (G, s) (K). As both α and γ are solutions of lDs (x) = a, it
follows that α−1 · γ ∈ (G, s) (K̂) = (G, s) (K). Thus clearly L = L1 .
Here are some alternative characterizations of differential Galois extensions:
Lemma 3.7. Let (G, s) be a K-large algebraic D-group over K, and L =
K(α) a differential field extension of K, where α is a solution of lDs (x) = a
(with a ∈ L(G)(K)). Then the following are equivalent:
(i) L and α satisfy Definition 3.5.
(ii) L is contained in some differential closure of K.
(iii) (G, s) (K) = (G, s) (L) and (G, s) is L-large.
Proof. (i) implies (ii) is given by Lemmas 3.4 and 3.6 and (ii) implies (iii)
follows from Remark 3.3.
(iii) implies (i). Assume L satisfies (iii). Then using Remark 3.3, we
have that (G, s) (L̂) = (G, s) (K). Now K̂ embeds in L̂ over K, hence by
Lemmas 3.4 and 3.6, there is a solution β ∈ G(L̂) of lDs (x) = a, such that
K[β] is a simple differential ring. Now α = β · g for some g ∈ (G, s) (K), so
clearly K[α] is also a simple differential ring.
Condition (iii) is the analogue of “no new constants” in the strongly
normal case.
Remark 3.8. Suppose K is algebraically closed. Then the differential Galois extensions of K in the sense of Definition 3.5 coincide with the generalized strongly normal extensions of K in the sense of [8]. In particular,
L is a strongly normal extension of K (in the sense of Kolchin) just if L
is a differential Galois extension of K for an equation lDs (x) = a, on an
algebraic D-group (G, s), where G is defined over CK and s = s0 .
Proof. This follows from Proposition 3.4 of [8].
We now point out that given L as above, Aut∂ (L/K) has the structure of a
differential algebraic group (over K) in two different ways. One corresponds
to the usual differential Galois group in the linear case, and is simply of the
354
ANAND PILLAY
form (H, s) (K), where H is an algebraic D-subgroup of (G, s). The other,
corresponding to the “intrinsic” Galois group introduced by Katz, is of the
form (H1 , s1 ) (L) for s1 another algebraic D-group structure on G (defined
over K), and H1 a D-subgroup of (G, s1 ) defined over K. The algebraic
D-groups (H, s) and (H1 , s1 ) will be isomorphic, but not necessarily over
K, unless they are commutative. (But of course if K is algebraically closed,
H and H1 will be isomorphic over K as algebraic groups.)
So fix L = K(α) as in Definition 3.5. By 3.7 we have L < K̂ for some
fixed copy of the differential closure of K. Aut∂ (L/K) denotes the group
of differential field automorphisms of L over K. Note that for any σ ∈
Aut∂ (L/K), σ(α) is also a solution of lDs = a, hence σ(α) = b · c(σ) for a
unique c(σ) ∈ (G, s) (L) = (G, s) (K) (= (G, s) (K̂)).
Lemma 3.9. The map c above is a group isomorphism between Aut∂ (L/K)
and (H, s) (K) for some algebraic D-subgroup H of G defined over K.
Proof. As an automorphism σ of L over K is determined by its action on α,
clearly the map is a group isomorphism with its image. So all that we need
is that the image is of the required form. The model-theoretic proof of this
goes through showing that the image is a definable subgroup of (G, s) (K̂),
and then using quantifier-elimination for differentially closed fields. We will
give an algebraic proof, after first discussing the second incarnation of the
Galois group.
First, the equation lDs (x) = a equips G with another structure of an
algebraic D-variety (but not in general an algebraic D-group). Let ra be
the right invariant vector field on G determined by a on G. So s + ra is
a K-rational section of τ (G) → G, which we denote by s . Note that the
equation lDs (x) = a on G is equivalent to the equation ∂(x) = s (x).
Now G acts on itself by left translation. Let S < G be the intersection
of the stabilizers of the algebraic D-subvarieties of the algebraic D-variety
(G, s ) that are defined over K: namely (working in some algebraically closed
field containing K), S = {g ∈ G : g · X = X for all D-subvarieties X of
(G, s ) defined over K}. S is an algebraic subgroup of G defined over K. S is
precisely the analogue in our context of the intrinsic differential Galois group
introduced by Katz in the Picard–Vessiot case and discussed by Bertrand
in [1].
In fact G can be naturally equipped with another structure of an algebraic
D-group, (G, s1 ). Define the section s1 : G → τ (G) by
s1 (g) = s(g) + ra (g) − la (g).
Lemma 3.10. (G, s1 ) is an algebraic D-group, and the action of G on itself by left multiplication is an action of the algebraic D-group (G, s1 ) on
the algebraic D-variety (G, s ). Moreover S is an algebraic D-subgroup of
(G, s1 ).
ALGEBRAIC D-GROUPS
355
Proof. An easy computation. Note that it follows that (G, s1 ) (U) acts on
(G, s ) (U).
Let Z be the set of solutions of lDs (x) = a in G(L). Note that Z is
precisely (G, s ) (L). In any case, Z is a principal homogeneous space for
(G, s) (K) (acting on the right). As (G, s) (K) = (G, s) (K̂), and L < K̂,
X is also the solution set of lDs (x) = a in G(K̂). Clearly Aut ∂ (L/K) acts
on Z and any σ ∈ Aut ∂ (L/K) is determined by its action on Z. In fact,
as L = K(β) for any β ∈ Z, σ ∈ Aut ∂ (L/K) is determined by any pair
(β, σ(β)) for β ∈ Z.
Lemma 3.11. The action of Aut ∂ (L/K) on Z is isomorphic to the action
(by left multiplication) of (S, s1 ) (L) on Z: The isomorphism d say takes σ
to σ(β) · β −1 for some (any) β ∈ Z.
Proof. We know that (S, s1 ) acts on Z by left multiplication (by Lemma 3.10). Suppose σ ∈ Aut ∂ (L/K). Let d(σ) ∈ G(L) be such that σ(α) =
d(σ) · α. As σ(α) ∈ Z, we have by Lemma 3.10 that d(σ) ∈ (G, s1 ) (L). As
any β ∈ Z is of the form α · c for c ∈ (G, s) (L) = (G, s) (K) and σ fixes K
pointwise, we see that
(∗)
σ(β) = d(σ) · β for all β ∈ Z.
We only have to see that d(σ) ∈ S. Let X be any D-subvariety of (G, s )
defined over K. As K̂ is differentially closed and Z = (G, s ) (K̂), it follows
that Z ∩ X is Zariski-dense in X. For β ∈ Z ∩ X, σ(β) ∈ Z ∩ X too. By (∗)
and Zariski-denseness, d(σ) · X = X. Thus d(σ) ∈ S.
Conversely, suppose that g ∈ (S, s1 ) (L). Then g · α ∈ Z. Let X be
the algebraic variety defined over K whose K-generic point is α. Then,
X is a D-subvariety of (G, s ). Thus g · X = X and so g · α ∈ X, and
∂(g · α) = s (g · α). Hence g · α is a differential specialization of α over
K. By simplicity of K[α], α and g · α satisfy exactly the same differential
polynomial equations over K. As both α and g · α generate L it follows
that there is an automorphism σ of L over K such that σ(α) = g · α. So
g = d(σ).
Conclusion of proof of Lemma 3.9. Let H be the image of S under conjugation by α−1 in G (g → α−1 · g · α). Then H is an algebraic D-subgroup
of (G, s). By 3.11 and what we already know about 3.9, the image of the
embedding c : Aut ∂ (L/K) → (G, s) (K) is precisely (H, s) (K). As the
latter is Zariski-dense in H, H is also defined over K.
Let us fix the isomorphism c between Aut ∂ (L/K) and (H, s) (K) =
(H, s) (K̂).
Lemma 3.12. There is a Galois correspondence between the set of differential fields in between K and L and the set of algebraic D-subgroups of
356
ANAND PILLAY
(H, s) defined over K (equivalently defined over K̂): given K < F < L the
corresponding group is HF the Zariski closure of the set of h ∈ (H, s) (K)
such that c−1 (h) is the identity on F .
Proof. This is Theorem 2.12 of [8] after making the translation between
definable subgroups and D-subgroups.
Let us complete this section with an example. We will consider the nonisotrivial algebraic D-group structure (G, s) on Gm × Ga discussed at the
end of Section 2, and exhibit a (natural) differential Galois extension K < L
whose differential Galois group is (G, s) (or rather (G, s) (K)). In fact there
will be an intermediary differential field K < F < L such that each of
K < F and F < K are Picard–Vessiot extensions, but K < L is not a
Picard–Vessiot extension.
Recall that s : G → T (G) is given by s(x, y) = (x, y, xy, 0), and lDs :
G → L(G) is lDs (x, y) = (∂(x)/x−y, ∂(y)), and so a logarithmic differential
equation on (G, s) has the form {∂(x)/x − y = a1 , ∂(y) = a2 }. If we restrict
our attention to equations where a1 = 0, we obtain equations of the form
∂(∂(x)/x) = a on Gm .
Also (G, s) can be identified with {x ∈ Gm : ∂(∂(x)/x)) = 0}.
Note that the embedding x → (x, 0) of Gm in G and surjection (x, y) → y
of G onto Ga induces an exact sequence
0 → (Gm , (s0 )Gm ) → (G, s) → (Ga , (s0 )Ga ) → 0
of algebraic D-groups, where s0 denotes the 0-sections for the corresponding
groups. The important fact is that (G, s) is not a product (as a D-group)
of the two groups, which is the reason that (G, s) is nonisotrivial.
We will take the ground field of constants to be C. Let K = C(ect : c ∈ C)
2
and L = K(t, et ). So L is a subfield of a fixed differential closure Ĉ = K̂ = L̂
of C.
Lemma 3.13. (G, s) is K-large.
Proof. It is enough to show that all solutions of ∂(∂(x)/x) = 0 in K̂ are in
K, that is all solutions of ∂(d)/d = c for c ∈ C that are in K̂ are in K.
But this is clear, because ect is one such such solution and the others are
obtained by multiplying by a constant.
Lemma 3.14. L is a differential Galois extension of K for the equation
∂(∂(x)) = 2 . The Galois group is (G, s) (K).
2
Proof. Note that L is generated over K as a differential field by et , which
is a solution of ∂(∂(x)) = 2. To show that the Galois group is as stated, it
is enough to show that tr.deg(L/K) = 2, which is well-known.
Note that K(t) is a Picard–Vessiot extension of K, and L is a Picard–
Vessiot extension of K(t), but L is not a Picard–Vessiot extension of K.
ALGEBRAIC D-GROUPS
357
4. Isomorphisms of algebraic D-groups and algebraic D-varieties
We will prove:
Proposition 4.1. Suppose (K, ∂) is an algebraically closed differential field,
(G, s) and (H, t) are connected algebraic D-groups over K and there is an
isomorphism f between (G, s) and (H, t) defined over some differential field
extension of K. Then there is such an isomorphism defined over a Picard–
Vessiot extension of K.
Here is a restatement of the theorem in the language of Kolchin’s constrained cohomology (see [5]).
Corollary 4.2. Suppose (K, ∂) has no proper Picard–Vessiot extensions.
Then for any connected ∂0 -group G defined over K, Hc1 (Aut ∂ (K̂/K), G(K̂))
is trivial.
In Proposition 3.11 of [9], Corollary 4.2 was stated in the special case that
H(K) = H(K̂) (which amounts to saying that the corresponding algebraic
D-group is K-large).
Kolchin’s differential tangent space and its properties (see Chapter 8 of
[5]) play an important role in the proof of Proposition 4.1. We will summarize the key properties (in the language of algebraic D-groups). Recall
first that if V is a finite-dimensional vector space over U, then a ∂-module
structure on V is an additive map DV : V → V such that DV (av) =
∂(a)v + aDV (v) for all a ∈ U and v ∈ V . V ∂ denotes {v ∈ V : DV (v) = 0},
a vector space over C with C-dimension the same as the U-dimension of V
Fact 4.3. Suppose (G, s) is a connected algebraic D-group defined over K,
and V = L(G) is its Lie algebra (tangent space at the identity). Then:
(i) s equips V with a canonical ∂-module structure DV , defined over K.
(ii) For any automorphism f of (G, s), dfe ∈ GL(V ) restricts to a C-linear
automorphism of V ∂ , and moreover f is determined by dfe |V ∂ .
Proof of Proposition 4.1. First (G, s) and (H, t) will be isomorphic over K̂.
Let f be such an isomorphism. Let c be a finite tuple from K̂ generating
the smallest field of definition of f . Write f as fc . Let (V, DV ) be as in Fact
4.3 for (G, s). Let d be a C-basis for V ∂ contained in K̂. Let L = Kc, d
be the differential field generated by c and d over K.
Claim I. L is a strongly normal extension of K.
Proof. As L < K̂, CL = CK . We have to show that for any automorphism
σ of the differential field U fixing K pointwise, σ(L) is contained in the
differential field generated by L and C. First σ(d) is another basis of V ∂ , so
with respect to the basis d we may write σ(d) as a n × n nonsingular matrix
B ∈ GLn (C). On the other hand fσ (c) = fc · h for a unique automorphism
358
ANAND PILLAY
h of (G, s). By Fact 4.3(ii), h is determined by dhe |V ∂ , which by in terms
of the basis d, is another nonsingular n × n matrix A over C. It follows
that (σ(c), σ(d)) is in the differential field generated by K, c, d, A and B. In
particular σ(L) ⊆ LC
.
Claim II. L is a Picard–Vessiot extension of K.
Proof. Let σ ∈ Aut ∂ (L/K), and write the matrices A, B (which will be
in GLn (CK )) as Aσ , Bσ . Then it is easy to check that σ → (Aσ , Bσ ) is
an embedding of Aut ∂ (L/K) into GLn (CK ) × GLn (CK ). From Kolchin’s
characterizations of Picard–Vessiot extensions, we get Claim II.
As f is defined over L, we have proved Proposition 4.1.
Finally let us give some additional results with a similar flavor. The first
is really just a remark and is related to the themes of Buium’s book [2].
Buium calls an algebraic D-variety (X, s) split if (X, s) is isomorphic (over
U) to a trivial algebraic D-variety, namely one of the form (Y, s0 ), where Y
is defined over C and s0 is the 0-section. He proves that any D-variety (X, s)
such that X is a projective variety is split. Moreover assuming Y defined
over algebraically closed K, then (Y, s) is split over some strongly normal
extension.
Remark 4.4. Let K be a differential field with algebraically closed constant
field. Suppose that (X, s) is an algebraic D-variety over K that is split.
Then (X, s) is split over some strongly normal extension K1 of K.
Proof. We can find an isomorphism f of (X, s) with some trivial (Y, s0 )
defined over K̂. Let again f = fc with c the smallest field of definition of f .
Let K1 = Kc
. Then, as K1 is contained in K̂ and CK̂ = CK , we have that
CK1 = CK . For any (differential) automorphism σ of U, Then fσ(c) = fc ◦ g
for some automorphism g of (Y, s0 ). But then g must be defined over C.
Hence σ(c) is rational over K(c)(C), so σ(K1 ) ⊆ K1 C
. This shows that K1
is a strongly normal extension of K.
The next result is a generalization of Proposition 4.1 in which the higherdimensional versions of differential tangent spaces from [10] enter the picture. We use freely the results from that paper.
Proposition 4.5. Suppose that K has algebraically closed constant field.
Let (X, s) and (Y, t) be algebraic D-varieties defined over K. Suppose that
a ∈ (X, s) (K), b ∈ (Y, t) (K) are nonsingular points on X, Y respectively,
and that there is some isomorphism f between (X, s) and (Y, t) such that
f (a) = b. Then there is such an isomorphism defined over a Picard–Vessiot
extension of K.
ALGEBRAIC D-GROUPS
359
Proof. For each m ≥ 1, let Vm the the U-vector space M/Mm+1 , where
M is the maximal ideal of the local ring of X at a. Vm is defined over
K. For any automorphism h of X such that h(a) = a, h induces a linear
automorphism hm say of Vm . Moreover if h is another automorphism of
X fixing a, then h = h if and only if hm = hm for all m. So far nothing
has been said about the D-variety structure. As a ∈ (X, s) , ∂ extends to a
derivation ∂ on the local ring of X at a that preserves M and all its powers.
This gives Vm the structure of a ∂ module (Vm , DVm ) over U, defined over
K (for all m). If h is an automorphism of the algebraic D-variety (X, s)
that fixes a then hm ∈ GL(Vm ) restricts to a C-linear automorphism h∂m
of GL(Vm∂ ) (where Vm∂ is the solution space of DVm = 0). Moreover hm is
determined by h∂m .
Now we may find an isomorphism f between (X, s) and (Y, t) such that
f (a) = b and f is defined over K̂. Let c generate the smallest field of
definition of f . So c is a finite tuple from K̂. Write f as fc . For any
differential automorphism σ of U that fixes K pointwise, σ(fc ) = fσ(c) is
also an isomorphism of (X, s) with (Y, t) taking a to b. Hence fσ(c) ◦ fc−1 is
an automorphism of (X, s) taking a to itself. Write hσ for this map.
Claim I. There is m such that for all σ, τ ∈ Aut ∂ (U/K), hσ = hτ if and
only if (hσm )∂ = (hτm )∂ .
Proof. This follows from compactness and the earlier remarks as the set of
hσ is a uniformly definable family of automorphisms of (X, s).
Now let m be as in Claim I and let d be a C-basis for (Vm )∂ .
Claim II. K1 = Kc, d
is a strongly normal extension of K.
Proof. Let σ ∈ Aut ∂ (U/K). As σ(d) is also a basis for (Vm )∂ , σ(d) ∈
Kd
C
. On the other hand, by virtue of d, (Vm )∂ can be identified with C r
for suitable r, and thus (hσm )∂ can be identified with an element of GLr (C).
By Claim I, it follows that σ(c) ∈ Kc, d
C
. Thus σ(K1 ) ⊆ K1 C
.
As in the proof of 4.1, we conclude that actually K1 is a Picard–Vessiot
extension of K. As c ∈ K1 , this gives the proposition.
Acknowledgements. This work was carried out while the author was an
Invited Professor at the University of Paris VI in June 2003, and while
visiting the University of Naples II in July 2003. I am indebted to Daniel
Bertrand at Paris VI for his hospitality, many helpful conversations, and
his interest in the work reported here. Thanks also to Zoe Chatzidakis of
University Paris VII-CNRS for her suggestions, and to Paola D’Aquino of
Naples II for her hospitality.
360
ANAND PILLAY
References
[1] D. Bertrand, Review of “Lectures on differential Galois theory”, by A. Magid, Bull.
Amer. Math. Soc., 33 (1996), 289–294.
[2] A. Buium, Differential Function Fields and Moduli of Algebraic Varieties, Lecture Notes in Mathematics, 1226, Springer, 1986, MR 0874111 (88e:14010),
Zbl 0613.12018.
[3] A. Buium, Differential Algebraic Groups of Finite Dimension, Lecture Notes in Mathematics, 1506, Springer, 1992, MR 1176753 (93i:12010), Zbl 0756.14028.
[4] E. Kolchin, Differential Algebra and Algebraic Groups, Pure and Applied Mathematics, 54, Academic Press, New York, 1973, MR 0568864 (58 #27929), Zbl 0264.12102.
[5] E. Kolchin, Differential Algebraic Groups, Pure and Applied Mathematics, 114, Academic Press, Orlando, 1985, MR 0776230 (87i:12016), Zbl 0556.12006.
[6] D. Marker, Manin kernels, in ‘Connections between Model Theory and Algebraic and
Analytic Geometry’, Quaderni di Matematica, Univ. Naples II, 2000, MR 1930680
(2003g:12009).
[7] D. Pierce and A. Pillay, A note on the axioms for differentially closed fields of
characteristic zero, J. Algebra, 204 (1998), 108–115, MR 1623945 (99g:12006),
Zbl 0922.12006.
[8] A. Pillay, Differential Galois theory I, Illinois J. Math., 42 (1998), 678–699,
MR 1649893 (99m:12009), Zbl 0916.03028.
[9] A. Pillay, Finite-dimensional differential algebraic groups and the Picard–Vessiot theory, in ‘Differential Galois Theory’, Banach Center Publications, 58, Polish Academy
of Sciences, 2002, 189–199, MR 1972454 (2004c:12008), Zbl 1036.12006.
[10] A. Pillay and M. Ziegler, Jet spaces of varieties over differential and difference fields,
Selecta Math., 9(4) (2003), 579–599, MR 2031753.
[11] B. Poizat, A Course in Model Theory. An Introduction to Contemporary Mathematical Logic, Universitext, Springer, New York, 2000, MR 1757487 (2001a:03072),
Zbl 0951.03002.
[12] M. van der Put and M. Singer, Galois Theory of Linear Differential Equations,
Grundlehren der Mathematischen Wissenschaften, 328, Springer, Berlin, 2003,
MR 1960772 (2004c:12010), Zbl 1036.12008.
Received January 20, 2004 and revised May 7, 2004. The author was supported by NSF
grant DMS 0300639 and NSF Focused Research Grant DMS 0100979.
Department of Mathematics
University of Illinois at Urbana-Champaign
Urbana, Illinois 61801-2975
E-mail address: [email protected]
This paper is available via http://www.pacjmath.org/2004/216-2-8.html.